Employing ensemble reasoning to support clinical decision-making

Abstract

The combination of multiple forms of reasoning in conjunction is often used by physicians making clinical decisions. The Select and Test Model, which involves various forms of reasoning including abstraction, deduction, and abduction, is an epistemological framework that represents how reasoning can be employed in a clinical setting. While the subset of Artificial Intelligence (AI) research that is referred to as hybrid reasoning has involved the combination of various forms of reasoning, such as assertional ABox and terminological TBox reasoning, symbolic and statistical reasoning, or static and temporal reasoning, surprisingly little analysis has been conducted on the interaction between deductive and abductive reasoning. Since the term hybrid reasoning has various definitions throughout the literature, we instead refer to the integration of multiple types of reasoning as ensemble reasoning. Several challenges related to ensemble reasoning and their associated research questions naturally arise. Are the recommendations derived using our approach clinically relevant for use cases involving clinical reasoning? Are the factors used to arrive at recommendations clinically relevant? Is it clinically important or advantageous to reveal the recommendation rationale? When considering the interaction of multiple types of reasoners, does the order of reasoning make an impact on the results of the reasoning or the time it takes to obtain the results? Does an iterative approach versus an alternating approach make an impact on the results or the time it takes to obtain the results? Why might one approach be preferred over the other? Is it more efficient to split abducibles into subsets, run the abductive reasoner on each subset, and then combine the results than running an abductive reasoner with all abducibles specified? How can all interrelated abducibles be determined? We attempt to address the questions above and the soundness of our proposed approach by laying a theoretical foundation built on logical analyses. This foundation is then leveraged to design an ensemble reasoning system that we apply to several use cases related to the diagnosis and treatment of type 2 diabetes. We present the recommendations and justifications derived from our system to clinicians in order to evaluate the clinical relevance and usefulness of our approach. By implementing an evidence-based clinical decision-support system that employs ensemble reasoning, we take a step toward advancing the state-of-the-art of AI.